Deriving MFCC Parameters from the Dynamic Spectrum for Robust Speech Recognition

Abstract

State-of-the-art automatic speech recognition systems typically adopt the feature set containing mel-frequency cepstral coefficients (MFCC) and their time derivatives. The noise vulnerability of MFCC significantly degrades the recognition performance of such systems in noisy conditions. This paper describes a noise-robust feature extraction method. A set of new MFCC features is derived from the dynamic spectrum instead of the static spectrum as in the conventional MFCC feature extraction. It is shown that the dynamic spectrum preserves the spectral envelope information and, at the same time, is more noise resistant than the static spectrum. Experiments on Aurora 2 database show the noise robustness of the proposed features and it is preferable to replace MFCC with the new features in the state-of-the-art feature set.